What the Data for This Category Looks Like
The data sources for quality traceability in semiconductor wafer fabrication include MES systems, equipment FDC data, outputs from defect detection and metrology tools, wafer maps, process recipe libraries and experimental data. Daily documents cover process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect atlases and yield reports. Data update frequency is per wafer batch, real-time or daily. Fields include nanoscale process parameters, wafer batch numbers, defect coordinates and equipment operating parameters, with large variations in document length.
Constraints These Characteristics Impose on Vector Models and Indexes
The ultra-long process chain and multi-source heterogeneous data characteristics require vector models to support semantic association retrieval across processes and equipment. Real-time and high-frequency updated data requires adaptive incremental indexing mechanisms to avoid performance loss from full reconstruction. Fine analysis of nanoscale process parameters and defect spatial distribution requires sufficiently small chunking granularity to preserve local features. The limited rework window requires retrieval latency to be controlled within a reasonable range to support rapid decision-making.
How to Set Configurations
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the description length of a single process segment on a wafer process card, avoiding semantic breaks across processes |
| Overlap Length | 150–200 characters | Covers handover parameters between processes, preventing loss of associated logic in chunks |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
defaultToken | 2048 token | Adapts to the text length of a single exception handling sheet, meeting basic semantic recall requirements |
| Similarity Threshold | Calibrated based on actual measurements | Matches the feature matching accuracy requirements of defect atlases |
| Single Search Limit | Top 10 results | Balances recall efficiency and information completeness for cross-process root cause analysis |
Three Common Misconfigurations
- A 404 error is returned when configuring the Doubao embedding model API, caused by failing to correctly fill in the exclusive API domain assigned by the platform, leading to incorrect request routing.
- No improvement in recall rate after batch re-embedding of the knowledge base, caused by failing to adjust chunk length for nanoscale parameters in wafer data, leading to failed semantic chunking.
- Indexes corresponding to manually inserted knowledge base entries disappear several hours later, caused by failing to enable persistent storage for incremental indexing, leading to automatic cleanup of temporary indexes.
How to Confirm Proper Configuration
- Import a single wafer defect atlas document, and verify that the chunked text fragments cover complete defect descriptions and corresponding process information.
- Initiate a query containing process parameters and defect types, and confirm that the returned search results include associated data across equipment and processes.
- Check the FastGPT v4.17.0 system logs, and confirm that all embedding request status codes are 200, with no timeouts or routing errors.
- Manually trigger a batch re-embedding task, and check whether the vector index update progress bar for the knowledge base advances normally.
The values given are common starting points and should be measured against the reader's own samples.